CtxSift came from a problem I kept hitting while using coding agents: long-running sessions were burning through usage windows faster than expected, especially when the agent had to look at huge terminal outputs, test logs, stack traces, build errors, and CI noise. The annoying part was not just the raw output size. After context compaction, agents often start re-reading files and re-running commands to reconstruct what already happened. So even when I compressed command output, some of the token savings were getting paid back later as a recollection tax. I wanted something smaller than a full memory system, and simpler than running another MCP server or knowledge graph. So CtxSift became a lightweight command-output layer for coding agents: - `compress` keeps only the useful debugging signal from noisy command output - `recall` lets the agent recover previous command evidence from a local, workspace-scoped cache The goal is simple: help coding agents waste fewer tokens, repeat fewer commands, and recover context faster during long engineering sessions. CtxSift can use local or remote models, works with CPU/GPU setups, and is designed to fit into existing coding-agent workflows rather than becoming another heavy system to operate. I’d love feedback from people using Codex, Claude Code, Cursor, Aider, Copilot, OpenCode, Roo, Cline, Windsurf, Gemini CLI, Qwen Code, or any other agent-heavy workflow.

CtxSift

Save tokens and extend your coding sessions.

Save tokens and extend your coding sessions.